miRNA targets and target interaction networks

Recent studies showed a high reliability of miRNA targets predicted by TargetScan [7]. Therefore we selected the targets for all human miRNAs listed in the TargetScan database. We obtained a set of 677 miRNAs and 18,880 unique target proteins. The resulting miRNA-protein network contained 224,316 interactions. To predict miRNA targets based on PAR-CLIP data, the crosslink-centered regions (CCRs) from combined AGO-PAR-CLIP libraries [13] were used. Target site prediction for all CCRs was done with the program RNAhybrid [14] with the default parameters. From the resulting list we filtered all predictions with a p-value below 0.02 and an energy score below the 25% quantlile. This resulted in a final miRNA- mRNA list of 50,160 predicted interactions.

Association of protein complexes with miRNA target sets - test for statistical significance

We used the Fisher's exact test for assigning the significance of the association with protein complexes for each miRNA target set. The hypergeometric P-value is given as the probability under which we could expect at least N _c miRNA targets by chance in a protein complex, if we randomly select N _t (total number of miRNA targets) proteins out of the total set of proteins N consisting of all miRNA targets N _T and all proteins in complexes N _C . P-values were corrected for multiple testing of 677 miRNAs using the Holm-Bonferroni correction method. We assigned the association of complexes and miRNA clusters by using the union of targets from all miRNAs within one cluster. Here, we tested for significant overlaps of these unified sets between the components of a complex in the same way as for single miRNA target sets.

Enrichment of biological processes

In order to test for significant enrichment of biological functions based on Gene Ontology (GO) [15] and KEGG [16] pathways within the set of targets in protein complexes, the R package GOstats [17] was used. A set of targeted components of 722 targeted protein complexes was extracted and compared to a set of proteins which consisted of all components of these complexes.

Comparison of fold change distributions

We used fold change measurements after over-expression of selected miRNAs from recent proteomics studies [6, 7]. We selected for every of these miRNAs the protein complexes consisting of at least one of its targets. A set of components of these protein complexes was built. Within this set, we compared the fold changes of components that are targets of the specific miRNA with the fold changes of the non-target components. This was done by performing a one sided Kolmogorov-Smirnov test for each of the miRNAs that were investigated in the proteomics studies.

Cell culture

PANC-1 cells were purchased from ATCC (Manassas, VA, USA). PANC-1 stable clones for miR-141 or miR-200c were obtained with sequence verified pRetroSuper-miRNA plasmids. Cell lines were cultivated under standard conditions in DMEM + 10% fetal bovine serum + 2 μg/ml puromycin. For transient knock down PANC-1 were transfected with siRNA targeting ZEB1 (r(aga uga uga aug cga guc g)d(TT)), CtBP2 (1: r(cuuuggauucagcgucaua)d(TT), 2: r(cuuuguaacugauucugga)d(TT)) or GFP (r(gcu acc ugu ucc aug gcc a)d(TT). All transfections and reporter assays were performed as described previously [18].

Specific assay for miRNA modulation

RNA from cultured cells was extracted using the mirVana™ miRNA Isolation Kit (Ambion, Austin, TX, USA). mRNA expression values were measured in triplicate using the Roche LightCycler 480 and normalized to b-actin expression as a housekeeping control. Expression values were calculated according to ref.[19].

Immunoblots

were performed using modified standard protocols. In brief, whole cell extracts were made of the cells in Triple Lysis Buffer [50 mM Tris-HCl pH8, 150 mM NaCl, 0,02% (w/v) NaN₃, 0,5% (w/v) NaDeoxycholate, 0,1% SDS, 1% (v/v) NP40]. Extracts (10 μg/lane) were separated on a 10% SDS-polyacrylamide gel, blotted onto a PVDF membrane, and incubated with the indicated primary antibodies diluted in blocking buffer (5% nonfat dry milk) over night at 4°C. After washing and incubation with peroxidase-coupled species-specific secondary antibodies, the signal was developed using SuperSignal West PICO Chemiluminescent Substrate (Perbio Science, Bonn, Germany) according to manufacturer's protocol. CtBP2, CDYL, RCOR3, β-actin and ZEB1 were immunodetected with the following primary antibodies: anti-CtBP2 mouse monoclonal antibody (1:8.000, BD Transduction Laboratories™, Franklin Lakes, NJ, USA), anti-CDYL rabbit polyclonal antibody (1:500, Abcam, Cambridge, UK), anti-RCOR3 rabbit polyclonal antibody (1:1000Abcam, Cambridge, UK) anti-β-actin mouse monoclonal antibody (1:5.000, Sigma-Aldrich Chemie GmbH, Munich, Germany). The anti-ZEB1 rabbit polyclonal antibody (1:20.000) was a gift of D.S. Darling, University of Louisville, Louisville, KY, USA.

Functional spectrum of miRNA-regulated protein complexes

We next analyzed the spectrum of functions covered by our set of miRNA-regulated protein complexes. We identified the biological processes (Gene ontology categories [15]) and pathways representing the molecular interactions and reaction networks (KEGG [16]), which are enriched within the total set of 810 miRNA-targeted components of the protein complexes (Additional file 3, Table S3 and Additional file 4, Table S4 online). In all, as shown in Figure 1a, the miRNA-regulated protein complexes are mainly involved in regulation of RNA metabolic process, regulation of transcription and chromatin modification. Conversely, house-keeping functions, such as translational elongation and ATP synthesis coupled electron transport are underrepresented. The results confirm earlier investigations [21] showing that miRNAs less frequently target genes involved in essential cellular processes. Interestingly, there is an overrepresentation of genes involved in the G1 phase of mitotic cell cycle, while genes that are involved in the S phase and the M/G1 transition of mitotic cell cycle are underrepresented. Experimental evidence has already been reported for the regulation of signal transduction in several metazoan species [22–26] and the cell cycle [27, 28] by miRNAs. The regulation of the cell cycle by miRNAs is further supported by strong correlations of miRNA over-expression with different types of cancer [29].

These observations correspond with the overrepresentation of targeted genes contained in pathways from KEGG (see Figure 1b). A high overrepresentation of genes could be observed in "Pathways in cancer". Also many signaling pathways are overrepresented, namely Wnt signaling, TGF-beta signaling, Insulin signaling, Notch signaling, ErbB signaling, MAPK signaling, T and B cell receptor signaling and Chemokine signaling. Genes involved in house-keeping functions were underrepresented also in KEGG pathways, namely RNA polymerase, RNA transport, Proteasome, Oxidative phosphorylation and Ribosome.

Validating predicted miRNA targets in protein complexes

Two recent proteomics studies measured the changes in synthesis of proteins in response to miRNA over-expression or knockdown on a genome-wide scale for selected miRNAs [6, 7]. We incorporated the data of these studies in order to validate our predictions. To determine the impact of protein downregulation by miRNAs, which have targets in protein complexes, the level of downregulation of targeted components and non-targeted components was compared. We considered both significantly and insignificantly regulated complexes, since the amount of significantly regulated complexes for the examined miRNAs in the proteomics study is too low to provide statistical significance. The negative fold changes of the targeted components were significantly higher than the negative fold changes of the non-targeted components (see Table 3 and Figure 2) for every analyzed miRNA. For example, our data showed that the LARC (LCR-associated remodelling) complex [30] has two (out of 19) components, which are computationally predicted targets of let-7. These two components, namely DPF2 (Zinc finger protein ubi-d4) and SMARCC1 (SWI/SNF-related matrix-associated actin-dependent regulator of chromatin subfamily C member 1) were modestly down-regulated (fold changes of -0.38, and -0.2, respectively), when let-7b was over-expressed in HeLa cells [7]. LARC binds to the DNase hypersensitive 2 site in the human β-globin locus control region (LCR) and transactivates β-like globin genes [30]. By simultaneously down-regulating two components of the LARC complex, let-7b might contribute to the overall transcriptional repression of the human β-globin locus.

miRNA	P-value
hsa-let-7b	1,5E-05
hsa-miR-1	5,1E-12
hsa-miR-155	5,3E-04
hsa-miR-16	1,8E-08
hsa-miR-30b	1,6E-03

The P-value was calculated by a one-sided Kolmogorov-Smirnov test. It was used to compare the fold change distributions of complex components that are miRNA targets and non-target ones.

PAR-CLIP (Photoactivatable-Ribonucleoside-Enhanced Crosslinking and Immuno-precipitation) is a powerful tool to detect segments of RNA bound by RNA-binding proteins (RBPs) and ribonucleoprotein complexes (RNPs). We corroborated the miRNA target sites identified by PAR-CLIP [13] with the proteomics data [6, 7]. 55% of the proteins with miRNA targets sites predicted based on PAR-CLIP data were moderately down-regulated (log2-fold change < -0.1). 413 protein complexes contained miRNA target sites in at least two subunits (Additional file 5, Table S5 online). Interestingly, of the 5,185 unique proteins with miRNA target sites identified based on PAR-CLIP data, 607 (12%) are members of protein complexes (with at least two distinct targets of one miRNA in the same protein complex). For comparison, the manually curated collection of human protein complexes in the CORUM database covers 2,780 unique proteins (2% of UniProt proteins). This implies miRNA targets identified from PAR-CLIP data are more likely to be in a protein complex from the CORUM database (12%) as compared to proteins in general (2%). While miRNAs frequently target multiple genes with isolated functions, these independent data, though only by a simple estimate, suggest that there is also a significant proportion of miRNA targets, which are distinct members of protein complexes (hypergeometic P-value 1.23e-11).

Protein complexes and miRNA expression

We next tested whether miRNAs, which target different components of the same protein complex, are more likely to be co-expressed. The average expression correlation (Co-expression as calculated by Pearson correlation coefficients, hereafter termed PC values) of miRNAs was examined based on pairwise correlation calculations of miRNA expression profiles obtained for 26 different organ systems and cell types [31]. To test for statistical significance, we combined all pairwise PC values obtained from the sets of miRNAs which significantly target the same complex. These PC values were then compared to all other pairwise PC values that were present in the data set from [31]. We performed a one-sided Kolmogorov-Smirnov (KS) test for the two PC value distributions and obtained a significantly (P-value 6.106e-24) higher co-expression within the sets of miRNAs that target the same complex. Since we are interested in coexpression of miRNAs that are not in one transcription unit, we also tested for increased correlation only for miRNAs of different transcription units. Only a few (3.3%) of the correlated miRNAs were actually contained in one transcription unit. Therefore, the result remains highly significant (P-value 2.11e-18). Another bias of our results might occur due to fact that all miRNAs from one family must target the same complex since they target the same set of mRNA. We compared only miRNAs within one complex that belong to different families. The KS test resulted in a P-value of 0.0058. Taken together, our statistical test indicates that miRNAs targeting different components of a protein complex are significantly co-expressed. The average Pearson correlations of miRNAs that simultaneously target a specific complex can be found in Additional file 6, Table S6 online1).

Protein complex networks co-ordinately regulated by clusters of miRNAs

We systematically characterized the protein complex networks, which are simultaneously regulated by clustered miRNAs in 154 transcription units gained from miRBase [1]. The interconnectivity of the target sets of the miRNA gene clusters was first assessed as follows: the number of protein-protein interactions between the target sets of each pair of miRNAs in the cluster was counted, and these values were compared to 1,000 randomly sampled sets of miRNAs. To avoid miRNA target prediction bias arising from redundant prediction of clustered miRNA family members, only targets of one family member were counted within each cluster. The statistical analysis revealed 35 clusters, whose targets are significantly interconnected in the protein-protein interaction network (P-value < 0.05, permutation test, 1,000 samples, Table 1). Comparing the observed number of interactions (Figure 3b) with the corresponding distributions of randomly sampled sets of miRNAs provides a strong indication that a significant fraction of miRNAs in clusters might co-ordinately regulate targets (P-Value < 0.02, Wilcoxon signed rank test, Additional file 7, Table S7 online). In order to support this finding, we also applied Fisher's exact test to test if the global number of target interactions from miRNA clusters is higher than expected by chance. This test resulted in a P-value < 2e-16.

CtBP/ZEB complex regulated by the miR-141-200c cluster

Very recent reports have shown that the miR-200 family regulates epithelial to mesenchymal transition (EMT) by targeting the transcriptional repressor zinc-finger E-box binding homebox 1 (ZEB1) and ZEB2[4, 32–35]. During EMT, the miR-141-200c cluster and the tumor invasion suppressor gene E-cadherin are downregulated by ZEB1/2[35]. ZEB1 and ZEB2 repress transcription through interaction with corepressor CtBP (C-terminal binding protein) [36]. Interestingly, several essential components of the CtBP/ZEB complex, namely ZEB1/2, CtBP2, RCOR3 (REST corepressor 3) and CDYL (Chromodomain Y-like protein), are predicted targets of the miR-141-200c cluster. CtBP2 has one miR-141 target site and one miR-200c target site, while ZEB1 and CDYL have two miR-200c target sites. RCOR3 has one miR-141 target site. The CtBP/ZEB complex mediates the transcriptional repression of its target genes by binding to their promotors and altering the histone modification [37].

We showed that overexpression of miR-141 and miR-200c led to reduced expression of CtBP2 and ZEB1 in human pancreatic carcinoma (PANC-1) cells (Figure 4a). Luciferase reporter assay showed reduced activity of the CtBP2 and ZEB1 3'UTR-luciferase reporters with increased levels of miR-141 and miR-200c (Additional file 8, Figure S1 online). These results are also confirmed on protein level by immunoblots (Figure 4b). In order to rule out the possibility that the stability of ZEB1 and CtBP2 are dependent on each other, we separately knocked down ZEB1 and CtBP2 by siRNAs in PANC-1 cells and observed no change in protein levels of the respective complex partner (Figure 4c). Although the expression of CDYL and RCOR3 is less obviously affected by overexpression of miR-141 and miR-200c in PANC-1 cells as compared to CtBP2 and ZEB1 (data not shown), we observed a downregulation of CDYL and RCOR3 on the protein level, when miR-141 or miR-200c were transiently transfected in PANC-1 cells (Figure 4d), suggesting that CDYL and RCOR3 are also targets of the miR141-200c cluster. Together, these experiments demonstrate, for the first time, that CtBP2, CDYL and RCOR3 can be regulated by miR141-200c cluster post-transcriptionally. As the functional consequence of miRNA overexpression, the expression of E-cadherin mRNA is greatly upregulated (Figure 4a), indicating that the repression activity of CtBP/ZEB complex is compromised. The interaction between the miR-141-200c cluster and multiple components of the CtBP/ZEB complex suggests a coordinated regulation of the repression activity for the CtBP/ZEB complex. Intriguingly, the miR-141-200c cluster also targets β-catenin, which is a shared component of cell adhesion and Wnt signalling [38]. β-catenin is found in the plasma membrane, where it promotes cell adhesion by binding to E-cadherin, in the cytoplasm, where it is easily phosphorylated and degraded in the absence of a Wnt signal, and in the nucleus, where it binds to TCF transcription factors and induces the transcription of Wnt target genes. Most protein-interacting motifs of β-catenin overlap in such a way that its interactions with each of its protein partners are mutually exclusive [38]. Since the miR-141-200c cluster and E-cadherin are both downregulated during EMT, it is tempting to speculate that more β-catenin would be made available for participating in transactivating downstream genes, which may contribute to the progress of cancer [4].